Gasgoo Munich-On November 2, UBTECH officially launched Thinker-Sim, its embodied AI simulation platform.
Centered on a Real2Sim2Real loop, the platform spans real-scene digitization, scaled synthetic data generation, and robotic capability benchmarking. It takes aim at persistent industry bottlenecks: the high cost of real-world data collection, the complexity of building simulation environments, and lengthy algorithm validation cycles.
Gathering real-world data robot by robot hits hard limits on cost and speed, making simulation the primary engine for data at scale. Yet for simulation to matter, it first has to look the part — and that is exactly where Thinker-Sim steps in.
Image credit: UBTECH
On the scene-reconstruction front, UBTECH leans on its proprietary Thinker-GS high-fidelity algorithm and automated SimReady asset generation. The system maps real factories and commercial showrooms into 3D, preserving spatial layouts, environmental structures, and visual details. From there, it layers in foreground assets — robots, equipment, people — tailored to the task at hand, whether that' s training, data harvesting, or algorithm benchmarking, to quickly assemble interactive simulation environments.
When it comes to data synthesis, embodied AI models demand massive, high-quality, and diverse datasets. UBTECH answers this with a proprietary high-performance rendering engine, a massively parallel simulation architecture, and advanced synthesis algorithms. The result is a highly efficient data pipeline: rendering speeds climb 67.3%, a single 8-GPU node generates 390 hours of synthetic data daily, and custom SIL/HIL communication modules boost efficiency by more than 400% — gains that make large-scale parallel data harvesting possible.
For evaluation, the platform introduces a unified simulation benchmarking system covering perception, navigation, and manipulation tasks, allowing standardized validation and capability assessments across different robotic algorithms and models. By randomizing lighting, textures, and materials across multiple dimensions, it can recreate the unpredictability of real-world settings — a crucial step for testing algorithmic robustness and generalization.
Thinker-Sim also features standardized simulation templates for common training, synthesis, and evaluation workflows — all launchable with a single click. Visual workflows can be orchestrated without writing code, covering over 90% of target scenarios. Meanwhile, the proprietary Ulab-Agent lets users manage task pipelines using natural language.
Even so, synthetic data must eventually be validated on physical robots. Just how narrow the sim-to-real gap can become will ultimately hinge on how well the Real2Sim2Real loop performs in actual deployment.
